ApolloRaines/Mistral-Small-24B-Instruct-Jbliterated

Mistral-Small-24B-Instruct-Jbliterated: Refusal-Free LLM

This model, developed by Apollo Raines, is a 24 billion parameter instruction-tuned language model based on mistralai/Mistral-Small-24B-Instruct-2501. Its primary distinguishing feature is the surgical removal of refusal behaviors directly at the weight level, rather than through system prompts or inference-time patches. This means the model's weights no longer encode typical refusal responses like "I can't help with that."

Key Differentiators

• "Jbliteration" Method: Utilizes a multi-directional ablation technique (5 directions per layer across all 40 layers) to eliminate refusal behaviors while preserving core capabilities.
• Preserves Core Functionality: Null-space constraints and norm preservation were enabled during the modification process to ensure that the model's mathematical, coding, and reasoning abilities remain intact.
• Addresses Deeper Refusal Mechanisms: Unlike standard single-direction ablation, this method targets more subtle refusal tactics such as prompt reinterpretation, disclaimer injection, strategic omission, and safer framing, which are often found in other models.

Usage and Performance

This model serves as a drop-in replacement for its base model, offering a similar performance profile but without the inherent refusal mechanisms. It can be run on GPUs with limited memory by leveraging tools like DeepswapLLM, which streams layers across GPU, RAM, and disk, potentially running up to 4x faster than alternatives like AirLLM.

Ideal Use Cases

This model is particularly suited for applications where direct, unfiltered responses are required, and where the inherent refusal behaviors of standard instruction-tuned models are undesirable. Developers can integrate it using standard Hugging Face transformers library methods.